Scientists at St. Jude Children’s Research Hospital and the University of Wisconsin-Madison today share a generative algorithm for achieving both high genome and image resolution in spatial transcriptomics.

Co-first author Jiyuan Yang, PhD, St. Jude Department of Computational Biology and co-senior and corresponding author Jiyang Yu, PhD, St. Jude Department of Computational Biology interim chair look at spatial transcriptomics data.

Spatial transcriptomics is a cutting-edge technique that characterizes gene expression within sections of tissue, such as heart, skin or liver tissue. These snapshots provide insights into how spatial organization affects cellular functions across the spectrum of biology and disease. Up to now, researchers conducting spatial transcriptomics have had to choose between two options based on their needs: genome-wide coverage or single-cell resolution. To solve this tradeoff, scientists at St. Jude Children’s Research Hospital and the University of Wisconsin-Madison have created a computational tool that uses generative artificial intelligence (AI) to enhance the resolution of sequencing-based spatial transcriptomics without sacrificing gene coverage. The algorithm and its first findings were published today in Nature Methods.

“We’ve made the first generative algorithm that can predict spatial gene expression of whole transcriptomics at the single-cell level,” said co-senior and corresponding author Jiyang Yu, PhD, St. Jude Department of Computational Biology interim chair. “The key is to borrow information from single-cell RNA-seq atlas and histology imaging data with this generative modeling approach, giving full transcriptome coverage and single-cell resolution in spatial RNA-sequencing.”

The new computational tool, Spot imager with pseudo single-cell resolution histology (Spotiphy), employs a machine learning algorithm to improve conventional spatial transcriptomics technologies significantly. Those technologies look at predefined “spots” on a grid to capture gene expression. These are essentially the pixels that form the final gene expression image across the entire tissue section. Each spot typically contains multiple, often heterogeneous, cells, making them difficult to classify and analyze into single cells.

Schematic overview of Spotiphy and the matched datasets used for validation

a, Spotiphy workflow. Spotiphy requires three types of input data: scRNA-seq data, spot-level sequencing-based ST data and high-resolution histological images. The workflow consists of five major steps. (1) Reference construction. scRNA-seq data are used to select top marker genes and generate a signature reference for all cell types. (2) Segmentation. The histological image is used to count the number of nuclei in each spot. (3) Generative modeling. ST data are used to generate cell-type proportions and iscRNA data. (4) Imputation. Gaussian processes are used to impute cell-type proportions and iscRNA data for cells located outside the transcriptomic spots. (5) Transformation. The results of steps 3 and 4 are merged to produce a pseudo-single-cell-resolved whole-transcriptome image. b, Process of generating matched datasets for WT and AD mouse brain tissues. c, Uniform manifold approximation and projection (UMAP) of scRNA-seq reference data for the mouse brain. Twenty-two cell types from the Allen Brain Map Atlas (https://portal.brain-map.org/atlases-and-data/rnaseq) and seven cell types from our original data (red box) are included. Astrocytes and oligodendrocytes from both datasets are included for batch correction. Astrocyte_AI, astrocytes from the Allen Brain Map Atlas; Astrocyte_SJ, astrocytes from the original data; IT, interneuron; Oligo_AI, oligodendrocytes from the Allen Brain Map Atlas; Oligo_SJ, oligodendrocytes from the original data; CA, cornu ammonis; CT, corticothalamic; DG, dentate gyrus; Endo, endothelial; L, layer; Pvalb, parvalbumin; PT, pyramidal tract; SNCG, synuclein-γ; SST, somatostatin; SST Chodl, somatostatin and chondrolectin coexpressing; VIP, vasoactive intestinal polypeptide; NP, near-projecting; SUB, subiculum.

Spotiphy uses machine learning that extrapolates cell-type proportions and gene expression data to fill the space between “spots.” The program knows what to fill these gaps with because it was trained on vast spatial transcriptomics databases paired with histological images.

“Imagine a picture of a hand, but the middle is missing,” said co-senior author Junmin Peng, PhD, St. Jude Departments of Structural Biology and Developmental Neurobiology. “The algorithm has acquired general rules from its training so that it can impute the missing part of the picture — like reconstructing where and what the palm should look like, or in reality, the space between imaging spots.”

Spotiphy finds new insights from spatial transcriptomics data

Peng, who studies neurobiology, explained how single-cell spatial transcriptomics provides an opportunity to understand neurodegenerative diseases better, including Alzheimer’s, especially aberrant cells and their gene expression. However, “previous spatial transcriptomics data for Alzheimer’s have not truly resolved single-cell data,” Peng said. “Existing methods result in a low-resolution output, often clustering multiple cells into a single spot. We merged adjacent sections for RNA analysis and imaging one-to-one with Spotiphy, truly reaching a clean single-cell resolution with high gene coverage.

Applied to a mouse model of Alzheimer’s Disease, Spotiphy confirmed the findings of previous studies, validating the tool’s ability to pinpoint cell location and gene expression patterns accurately. This increased confidence in the technique. In addition, the researchers deployed Spotiphy in new experiments, finding that subsets of astrocytes, a common cell type in the central nervous system, are associated with specific brain regions. They also found that disease-associated microglia, a rare type of immune cells, had a greater presence in brains affected by Alzheimer’s, supporting previous findings that microglial dysfunction may be involved in the disease.

“The real power of this algorithm is its ability to distinguish subtle differences within the same cell type that previous technologies could not detect,” Yu said. “For example, identifying sub-populations of astrocytes located in certain areas of the brain.”

Findings were not limited to neurobiology. The scientists also showed that Spotiphy could be used to analyze other tissues, including cancer samples. The tool successfully identified different spatial domains and alterations in tumor-tumor microenvironment interactions.

“We saw the data reflected known breast cancer heterogeneity,” said co-first author Jiyuan Yang, PhD, St. Jude Department of Computational Biology. “We also invested significant efforts to generating the matched datasets for mouse brains, which we believe will be a valuable resource for the spatial omics community.”

“We’ve made a tool that can resolve single-cell spatial transcriptomics of any given tissue,” Yu said. “In other words, Spotiphy lets scientists see things that they could not see before.”

Availability – Spotiphy is freely available at https://github.com/jyyulab/Spotiphy.

Source – St. Jude Children’s Research Hospital